Update app.py
Browse files
app.py
CHANGED
|
@@ -1,28 +1,41 @@
|
|
| 1 |
import os
|
| 2 |
import shutil
|
| 3 |
import tempfile
|
|
|
|
| 4 |
import spaces
|
| 5 |
import gradio as gr
|
| 6 |
from fastapi import FastAPI, UploadFile, File, Form, HTTPException
|
| 7 |
from fastapi.responses import JSONResponse
|
| 8 |
-
from
|
| 9 |
|
| 10 |
-
# 1. 声明加载的模型
|
| 11 |
-
|
| 12 |
|
| 13 |
-
# 2.
|
| 14 |
-
#
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
def transcribe_core(audio_path: str):
|
| 17 |
-
#
|
| 18 |
-
|
| 19 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
-
|
| 22 |
-
text = "".join([segment.text for segment in segments])
|
| 23 |
-
return text
|
| 24 |
|
| 25 |
-
#
|
| 26 |
def gradio_predict(audio_path):
|
| 27 |
if audio_path is None:
|
| 28 |
return "请先上传音频或录音!"
|
|
@@ -36,10 +49,10 @@ demo = gr.Interface(
|
|
| 36 |
inputs=gr.Audio(sources=["microphone", "upload"], type="filepath", label="输入音频"),
|
| 37 |
outputs=gr.Textbox(label="识别出的文本"),
|
| 38 |
title="Whisper 语音识别 API 节点",
|
| 39 |
-
description="【A100 GPU 动态加速版】支持网页端测试,同时也支持 OpenAI 兼容的 /v1/audio/transcriptions 接口!"
|
| 40 |
)
|
| 41 |
|
| 42 |
-
#
|
| 43 |
app = demo.app
|
| 44 |
|
| 45 |
@app.post("/v1/audio/transcriptions")
|
|
@@ -62,6 +75,6 @@ async def transcribe_api(
|
|
| 62 |
|
| 63 |
return JSONResponse(content={"text": transcription_text})
|
| 64 |
|
| 65 |
-
#
|
| 66 |
if __name__ == "__main__":
|
| 67 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|
|
|
|
| 1 |
import os
|
| 2 |
import shutil
|
| 3 |
import tempfile
|
| 4 |
+
import torch
|
| 5 |
import spaces
|
| 6 |
import gradio as gr
|
| 7 |
from fastapi import FastAPI, UploadFile, File, Form, HTTPException
|
| 8 |
from fastapi.responses import JSONResponse
|
| 9 |
+
from transformers import pipeline
|
| 10 |
|
| 11 |
+
# 1. 声明加载的模型(Hugging Face 官方原生 Whisper Small,与 CTranslate2 版本一样精准)
|
| 12 |
+
MODEL_NAME = "openai/whisper-small"
|
| 13 |
|
| 14 |
+
# 2. 全局初始化 Pipeline,默认放在 CPU 上,防止启动报错
|
| 15 |
+
# generate_kwargs 指定中文识别
|
| 16 |
+
pipe = pipeline(
|
| 17 |
+
"automatic-speech-recognition",
|
| 18 |
+
model=MODEL_NAME,
|
| 19 |
+
chunk_length_s=30,
|
| 20 |
+
device="cpu"
|
| 21 |
+
)
|
| 22 |
+
|
| 23 |
+
# 3. 核心计算函数
|
| 24 |
+
@spaces.GPU(duration=60)
|
| 25 |
def transcribe_core(audio_path: str):
|
| 26 |
+
# 【核心安全操作】进入 GPU 节点后,动态将 Pipeline 的模型送入 CUDA 显存
|
| 27 |
+
pipe.model.to("cuda")
|
| 28 |
+
|
| 29 |
+
# 运行转录(开启 FP16 混合精度极速推理)
|
| 30 |
+
with torch.autocast("cuda"):
|
| 31 |
+
result = pipe(audio_path, generate_kwargs={"language": "chinese"})
|
| 32 |
+
|
| 33 |
+
# 转录完后立即将模型移回 CPU,完美符合 ZeroGPU 的释放规范
|
| 34 |
+
pipe.model.to("cpu")
|
| 35 |
|
| 36 |
+
return result["text"]
|
|
|
|
|
|
|
| 37 |
|
| 38 |
+
# 4. 创建 Gradio 界面
|
| 39 |
def gradio_predict(audio_path):
|
| 40 |
if audio_path is None:
|
| 41 |
return "请先上传音频或录音!"
|
|
|
|
| 49 |
inputs=gr.Audio(sources=["microphone", "upload"], type="filepath", label="输入音频"),
|
| 50 |
outputs=gr.Textbox(label="识别出的文本"),
|
| 51 |
title="Whisper 语音识别 API 节点",
|
| 52 |
+
description="【A100 GPU 动态加速版 - Transformers 官方兼容版】支持网页端测试,同时也支持 OpenAI 兼容的 /v1/audio/transcriptions 接口!"
|
| 53 |
)
|
| 54 |
|
| 55 |
+
# 5. 获取 FastAPI 实例并扩展 API 路由
|
| 56 |
app = demo.app
|
| 57 |
|
| 58 |
@app.post("/v1/audio/transcriptions")
|
|
|
|
| 75 |
|
| 76 |
return JSONResponse(content={"text": transcription_text})
|
| 77 |
|
| 78 |
+
# 6. 启动服务
|
| 79 |
if __name__ == "__main__":
|
| 80 |
demo.launch(server_name="0.0.0.0", server_port=7860)
|